𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗥𝗲𝗹𝗲𝗮𝘀𝗲𝘀 𝗠𝗔𝗜-𝗧𝗵𝗶𝗻𝗸𝗶𝗻𝗴-𝟭

Microsoft releases a new reasoning model. Microsoft calls it MAI-Thinking-1. It uses a mixture-of-experts design.

Check these scores:

  • 97% on AIME 2025 math.
  • 87.7% on LiveCodeBench v6 coding.
  • 52.8% on SWE-Bench Pro software engineering.

The model has 1 trillion total parameters. Only 35 billion stay active per token. This keeps costs low.

Microsoft uses a hill-climbing machine. This is a loop. It generates data. It trains. It tests. It repeats.

Microsoft trained it on 30 trillion human tokens. They avoided data from other AI models.

The process is the advantage. Microsoft creates a loop to improve models. This lets them compete with top AI labs.

Watch for the next version in 6 to 12 months. Check if the scores go up.

Source: https://dev.to/gentic_news/microsoft-unveils-mai-thinking-1-35b-active-1t-parameters-97-on-aime-2025-25db Optional learning community: https://t.me/GyaanSetuAi